Search papers, labs, and topics across Lattice.
MegaStyle introduces a data curation pipeline that leverages the style consistency of large text-to-image models to generate a large-scale (1.4M images) style dataset with intra-style consistency and inter-style diversity. They curate a prompt gallery of 170K style prompts and 400K content prompts, then combine them to generate the MegaStyle-1.4M dataset. Using this dataset, they train a style encoder (MegaStyle-Encoder) with style-supervised contrastive learning and a FLUX-based style transfer model (MegaStyle-FLUX), demonstrating the importance of dataset properties for style transfer tasks.
Style transfer just got a whole lot easier: MegaStyle's 1.4M image dataset, generated via text-to-image models, unlocks reliable style similarity measurement and generalizable style transfer.
In this paper, we introduce MegaStyle, a novel and scalable data curation pipeline that constructs an intra-style consistent, inter-style diverse and high-quality style dataset. We achieve this by leveraging the consistent text-to-image style mapping capability of current large generative models, which can generate images in the same style from a given style description. Building on this foundation, we curate a diverse and balanced prompt gallery with 170K style prompts and 400K content prompts, and generate a large-scale style dataset MegaStyle-1.4M via content-style prompt combinations. With MegaStyle-1.4M, we propose style-supervised contrastive learning to fine-tune a style encoder MegaStyle-Encoder for extracting expressive, style-specific representations, and we also train a FLUX-based style transfer model MegaStyle-FLUX. Extensive experiments demonstrate the importance of maintaining intra-style consistency, inter-style diversity and high-quality for style dataset, as well as the effectiveness of the proposed MegaStyle-1.4M. Moreover, when trained on MegaStyle-1.4M, MegaStyle-Encoder and MegaStyle-FLUX provide reliable style similarity measurement and generalizable style transfer, making a significant contribution to the style transfer community. More results are available at our project website https://jeoyal.github.io/MegaStyle/.
